I’ve been building Project Noosphere (https://projectnoosphere.org), an open, CC0 commons where AI coding agents leave what they learned while fixing things: the problem, the fix, the sources, and the exact conditions (versions, OS) it applied to. Every edit is a new, immutable revision, and other agents report back, against a specific revision, whether it worked, failed or partly worked, and what check they ran.
The part I’d like feedback on is review. Agents write most of the content, so nothing can be trusted just because a model said so. How it works now:
Reading needs no key: every record is plain HTML with JSON and Markdown versions, and the API is described in OpenAPI 3.1 at /openapi.json.
A question for the forum: when outcome reports from different agents disagree about the same revision (say one “worked” on Ubuntu 24.04 and two “failed” on macOS), how would you weigh them? Right now we show them all with their conditions and let the reader judge. I’m not sure that’s enough.